EvoLink Seedream 4.0 (Official)

Seedream is ByteDance's image model, and ByteDance's posture on it has never wavered: the flagship generation lines stay API-only while the components underneath ship open. You will not be downloading Seedream 4.0. EvoLink Seedream 4.0 (Official) is how it reaches your ComfyUI graph anyway - text-to-image, image-to-image, and image editing, with the standout feature being n: up to 15 images per call. That's the node's whole personality. It's the batch engine of the pack.

What it is

• prompt (required) - the description or edit instruction, Chinese native (this is a ByteDance model; Chinese prompts are its first language).
• size - auto, a ratio like 1:1 or 16:9 or 9:21, or explicit pixels like 2048x2048. Freeform string, not a dropdown.
• quality - 1K / 2K / 4K, default 2K.
• n - 1 to 15 images. The tooltip catches the one gotcha: the number is a cap - you also need to say in the prompt that you want multiple images, or the model may not actually split the batch. "生成 6 张" in the prompt + n=6 in the box.
• image1–image4 - reference images for img2img/editing, up to 14 total with the image_urls_text box. Local files upload via EvoLink's channel, auto-deleted after 72 hours.

Outputs: image (a tensor batch when n > 1 - connect Preview or Save Image), task_id, status_info, result_urls (24-hour expiry), response_json.

Install and key

Shared EvoLink install. ComfyUI Manager → search EvoLink → install → full restart (close the console, don't just refresh the browser). Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/deeplearning-goethe/comfyui-evolink

Dependency is just requests; no models, no GPU. Key from evolink.ai/dashboard/keys (sk-), pasted once into api_key - saved locally, blank after, never in a shared workflow JSON.

Why the 15-image cap is the point

Most cloud image nodes make you pay per render for "options." Seedream 4.0's n collapses that: one call, up to 15 takes, then you pick the winner and promote it to a 4K re-run. That's the most credit-efficient way to explore a design space that any of these closed image models offer, and it's the reason to reach for Seedream over a Nano Banana when you're doing variation hunting.

Watch the two traps. First, the prompt/cap coordination above - n alone won't force multi-image output if the prompt doesn't ask for it. Second, the standing cloud math: 15 images in one call is one job but it bills as a batch, and your prompts and reference images leave the machine. Moderation-blocked or failed jobs don't charge, and unchanged params cache - so when you find the winner, the re-run at 4K is the only new line item.